chore: timeSeries deep forecasters: correct autodiff training (GradientTape + IEngine) + GPU-residency seam#1841

Merged

ooples merged 13 commits into

feat/timeseries-tape-gpu

Jul 10, 2026

Commits on Jul 10, 2026

fix: fix NHiTS backward pass: train via GradientTape autodiff so the model actually learns

authored and
claude
committed
chore: batched GradientTape + IEngine forward (autodiff backward, GPU-dispatchable)

authored and
claude
committed
feat(timeseries): gPU-resident training seam for NBEATS via compiled fused plan

authored and
claude
committed
feat(timeseries): gPU-resident training seam for NHiTS via compiled fused plan

franklinic
and
claude
committed
feat(timeseries): gPU-resident training seam for Informer via compiled fused plan

franklinic
and
claude
committed
feat(tft): tape-train Temporal Fusion Transformer with batched IEngine ops

ooples
and
claude
committed
feat(autoformer): tape-based automatic-differentiation training (no hand-derived gradients)

ooples
and
claude
committed
feat(deepar): tape-based BPTT training + GPU-dispatchable forward

ooples
and
claude
committed
chore(timeseries): fix review comments — visibility, serialization, input validation, checkpoint

ooples
and
claude
committed
chore(timeseries): resident-path review fixes — holdout gate, dead-code, failed-step fallback

ooples
and
claude
committed
fix(nbeats): stop eager training divergence blowing up forecasts

ooples
and
claude
committed
fix(timeseries): resolve review comments — TFT quantile head, RandomHelper, facade, resident parity

ooples
and
claude
committed
Merge remote-tracking branch 'origin/master' into feat/timeseries-tape-gpu
ooples
committed